Skip to content

Pull requests: Avarok-Cybersecurity/atlas

New pull request New
22 Open 139 Closed
22 Open 139 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

test(harness): official-suite vision test + Holo-3.1/Ornith/27B matrix
#212 opened Jun 28, 2026 by rsafier • Draft
perf(loader): GPU NVFP4→BF16 dequant (was an 83M-iter CPU loop)
#211 opened Jun 28, 2026 by rsafier Loading…
feat(qwen35): native FP8 for Holo-3.1 (compressed-tensors float-quantized)
#210 opened Jun 28, 2026 by rsafier Loading…
fix(dflash): correctness fixes for K=γ verify path on Qwen3.6-27B NVFP4
#209 opened Jun 28, 2026 by Sujimoshi Loading…
2
grammar: budget-aware graceful close for structured outputs (#144) + #630 safety regression
#208 opened Jun 28, 2026 by tbraun96 Contributor Loading…
SBR M1: opt-in tail-pin SSM-snapshot eviction (8× warm-resume on deep agentic convs)
#207 opened Jun 28, 2026 by tbraun96 Contributor Loading…
8
fix(tool-parser): stop tool-call markup leaking into streamed content
#206 opened Jun 27, 2026 by rsafier • Draft
feat(holo): Holo-3.1-35B-A3B / Ornith GB10 enablement — runtime, MoE, GDN, attention, serving
#203 opened Jun 27, 2026 by rsafier Loading…
perf(vision): GEMM-based ViT attention + tensor-core block GEMMs + batched forward (~2× image TTFT, 7.2× ViT prefill)
#202 opened Jun 27, 2026 by rsafier Loading…
2
perf(gb10): int8 W4A8 prefill GEMM — faithful llama-MMQ port (faith2, 44.75/48.93 TFLOP/s)
#201 opened Jun 27, 2026 by tbraun96 Contributor Loading…
fix(qwen35): mixed-precision + per-channel FP8 expert loading (AgentWorld-35B, Ornith-1.0-35B-FP8)
#199 opened Jun 26, 2026 by tbraun96 Contributor Loading…
1
feat(kernel): lossless BF16 tensor-core dense-FFN prefill (w4a16_gemm_t_m128_bf16)
#198 opened Jun 24, 2026 by tbraun96 Contributor Loading…
chore(deps): Bump tower-http from 0.6.11 to 0.7.0
#190 opened Jun 22, 2026 by dependabot Bot Loading…
1
chore(deps): Bump the minor-updates group across 1 directory with 2 updates
#189 opened Jun 22, 2026 by dependabot Bot Loading…
1
chore(deps): Bump actions/checkout from 6 to 7 in the actions-all group
#188 opened Jun 22, 2026 by dependabot Bot Loading…
1
feat(strix): DP4A decode + tool multi-call — Atlas beats llama.cpp on Strix BFCL (89.22 > 88.02)
#187 opened Jun 21, 2026 by tbraun96 Contributor • Draft
6
DeepSeek-V4-Flash (NVFP4) — clean port onto main [WIP]
#186 opened Jun 20, 2026 by tbraun96 Contributor • Draft
4 of 12 tasks
5
fix(qwen3.6): resolve agentic tool-calling degeneration — FP8 block-scaled prefill, thinking budget, tool-coerce
#177 opened Jun 17, 2026 by mission-deny-the-mission Loading…
3 tasks done
4
fix: support codex namespace tools
#171 opened Jun 15, 2026 by 0137 Loading…
4 of 6 tasks
2
chore(deps): Bump rand from 0.8.6 to 0.9.3
#168 opened Jun 15, 2026 by dependabot Bot Loading…
1
vLLM-parity generation control: remove server-side loop bandaids, enforce context window, add bounded tool-completion guard
#162 opened Jun 13, 2026 by tbraun96 Contributor Loading…
perf(dflash): fuse decode+verify into one weight sweep + reclaim Option-B KV blocks
#132 opened Jun 9, 2026 by rrstesiak Loading…
4
ProTip! no:milestone will show everything without a milestone.